Dformer: Dual-Path Transformers for Geometric and Appearance Features Reasoning in Diabetic Retinopathy Grading

Abstract

In fundus screening, the accuracy of automatic grading of Diabetic Retinopathy (DR) is crucial for the early detection and intervention of DR. Existing methods are inaccurate in the judgment of DR grading, especially the grade of mild Non-proliferative DR (NPDR). There are two reasons for this. On the one hand, the microaneurysm, which is the main judgment basis, is very small. Their false detection and missed detection rates are very high. On the other hand, the existing Convolutional Neural Network (CNN) structure lacks the ability of global reasoning and comprehensive analysis. To address these issues, this paper proposes a transformer-based dual-path reasoning network that can infer from geometric and appearance features based on the preliminary clues discovered by the detection network. We also design exchange connections between the two paths to better integrate the learned weights. The proposed method can better identify early mild NPDR, thereby improving the accuracy of the overall DR grading. It was experimentally verified on the public DDR dataset and Messidor dataset, and the accuracy of early mild NPDR grading increased by 18.16%, and the overall DR grading accuracy increased by 1.93%. Ablation experiments and visualization analysis show that the proposed method can focus on lesions in images and effectively infer the correct DR level based on them.